kazalbrur/bangla-embed-e5-small

Bangla-Embed-E5

A compact 118M Bengali sentence encoder for retrieval, semantic textual similarity, and classification. Rebased from intfloat/multilingual-e5-small by three-stage distillation from a BGE-M3 teacher (cross-lingual distillation → supervised contrastive fine-tuning → NLI polish).

On the official MTEB(Indic) Bengali subset it scores mean 0.666 and is a strong encoder in the ≤120M tier. For romanized/cross-script (Banglish) retrieval, use the companion kazalbrur/bangla-embed-e5-small-banglish.

Comparison — MTEB(Indic) Bengali subset

Main score per task type (mteb 2.12), single shared harness. This model is bolded.

Model	Params	Retr	Class	Bitext-G	Bitext-C	Rerank	STS	Clust	Mean
BGE-M3 (teacher)	568M	0.644	0.879	0.874	0.722	0.852	0.593	0.340	0.700
bangla-embed-e5-small-banglish	118M	0.791	0.832	0.826	0.654	0.840	0.596	0.349	0.698
mE5-large	560M	0.631	0.847	0.876	0.748	0.852	0.540	0.339	0.690
bangla-embed-e5-small (this model)	118M	0.572	0.848	0.832	0.668	0.840	0.554	0.349	0.666
mE5-small (base)	118M	0.535	0.832	0.848	0.699	0.835	0.538	0.310	0.656
LaBSE	109M	0.442	0.804	0.849	0.705	0.792	0.583	0.239	0.631
Vyakyarth	300M	0.629	0.762	0.853	0.576	0.767	0.423	0.343	0.622
pm-mpnet-base	278M	0.337	0.749	0.618	0.426	0.701	0.355	0.370	0.508

Retr=BelebeleRetrieval, Class=BengaliSentiment, Bitext-G/C=IN22 Gen/Conv, Rerank=WikipediaReranking, STS=IndicCrosslingualSTS, Clust=SIB200. The top-3 means span only 0.010 and are statistically indistinguishable under a 7-task paired bootstrap (treat sub-0.02 gaps as ties). This model improves over its mE5-small backbone (0.656 → 0.666) and leads the small-model field on Bangla retrieval.

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("kazalbrur/bangla-embed-e5-small")
q = model.encode(["ঢাকা কোথায়?"], prompt_name="query", normalize_embeddings=True)
d = model.encode(["ঢাকা বাংলাদেশের রাজধানী।"], prompt_name="passage", normalize_embeddings=True)
print((q @ d.T)[0, 0])

This is an E5-family model: prefix queries with query: and passages with passage: (the prompt_name argument above does this for you). Output dimension is 1024, L2-normalized.

Training & data

Three-stage curriculum (AdamW, cosine schedule, bf16). Distillation over ~18.7M EN–BN parallel pairs; supervised contrastive fine-tuning (MNR) on ~2.45M pairs (Bangla-native core + SWIM-IR + machine-translated MS MARCO-bn); NLI triplet polish (XNLI-bn). Released under the MIT license. Note that some training sources carry their own (in some cases non-commercial) terms — verify upstream data terms before commercial deployment.

Limitations

Standard Bangla only (regional varieties unrepresented); on full-corpus MIRACL-bn/Mr.TyDi-bn it trails larger models and the untuned backbone; all metrics are automatic. See the paper for full details.